Resource-Efficient Deep Neural Networks for Automotive Radar Interference Mitigation

Radar sensors are crucial for environment perception of driver assistance systems as well autonomous vehicles. With a rising number radar and the so far unregulated automotive frequency band, mutual interference is inevitable must be dealt with. Algorithms models operating on data required to run early processing steps specialized sensor hardware. This hardware typically has strict resource-constraints, i.e. low memory capacity computational power. Convolutional Neural Network (CNN)-based approaches denoising mitigation yield promising results in terms performance. Regarding however, CNNs exceed hardware's capacities by far. In this paper we investigate quantization techniques CNN-based signals. We analyze (i) weights (ii) activations different model architectures. reduced requirements storage during inference. compare with fixed learned bit-widths contrast two methodologies training quantized CNNs, straight-through gradient estimator distributions over discrete weights. illustrate importance structurally small real-valued base show that smallest models. achieve reduction around 80\% compared baseline. Due practical reasons, recommend use 8 bits activations, which require only 0.2 megabytes memory.